In this notebook, you will be putting your recommendation skills to use on real data from the IBM Watson Studio platform.
You may either submit your notebook through the workspace here, or you may work from your local machine and submit through the next page. Either way assure that your code passes the project RUBRIC. Please save regularly.
By following the table of contents, you will build out a number of different methods for making recommendations that can be used for different situations.
I. Exploratory Data Analysis
II. Rank Based Recommendations
III. User-User Based Collaborative Filtering
IV. Content Based Recommendations (EXTRA - NOT REQUIRED)
V. Matrix Factorization
VI. Extras & Concluding
At the end of the notebook, you will find directions for how to submit your work. Let's get started by importing the necessary libraries and reading in the data.
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import project_tests as t
import pickle
import plotly.graph_objects as go
%matplotlib inline
df = pd.read_csv('../data/user-item-interactions.csv')
df_content = pd.read_csv('../data/articles_community.csv')
del df['Unnamed: 0']
del df_content['Unnamed: 0']
# Show df to get an idea of the data
df.head()
# Show df_content to get an idea of the data
df_content.head()
Use the dictionary and cells below to provide some insight into the descriptive statistics of the data.
1. What is the distribution of how many articles a user interacts with in the dataset? Provide a visual and descriptive statistics to assist with giving a look at the number of times each user interacts with an article.
df.dtypes
df_content.dtypes
df.isnull().mean()
df_content.isnull().mean()
print (len(df['email'].unique()), len(df['article_id'].unique()), df.shape[0])
print (len(df_content[df_content['doc_status'] == 'Live']['article_id'].unique()))
dist_user_article_df = (df
.groupby(['email'])
.count()
.sort_values(['article_id'], ascending=False)['article_id']
.reset_index()
)
dist_user_article_df.columns = ['user', 'art_inter_num']
x = dist_user_article_df['user']
y = dist_user_article_df['art_inter_num']
# Use the hovertext kw argument for hover text
fig = go.Figure(data=[go.Bar(x=x, y=y,
hovertext=['27% market share', '24% market share', '19% market share'])])
# Customize aspect
fig.update_traces(marker_color='rgb(158,202,225)', marker_line_color='rgb(8,48,107)',
marker_line_width=1.5, opacity=0.6)
fig.update_layout(title_text='Distribution of User-Article Interaction')
fig.update_xaxes(title='User', showticklabels=False)
fig.update_yaxes(title='Article Interaction Number')
fig.show()
# Fill in the median and maximum number of user_article interactios below
median_val = dist_user_article_df['art_inter_num'].median() # 50% of individuals interact with ____ number of articles or fewer.
max_views_by_user = dist_user_article_df['art_inter_num'].max() # The maximum number of user-article interactions by any 1 user is ______.
2. Explore and remove duplicate articles from the df_content dataframe.
# Find and explore duplicate articles
df_content.head()
print (df_content['doc_status'].unique()) # check the category of doc_status
print (len(df_content['article_id'].unique()), df_content.shape[0]) # check the dunplicated number of articles in df
(df_content[df_content.duplicated(subset=['article_id'], keep=False) == True]
.sort_values(['article_id'],ascending=True))
# Remove any rows that have the same article_id - only keep the first
df_content = df_content.drop_duplicates(subset=['article_id'], keep='first')
print (len(df_content['article_id'].unique()), df_content.shape[0]) # check the dunplicated number of articles in df
3. Use the cells below to find:
a. The number of unique articles that have an interaction with a user.
b. The number of unique articles in the dataset (whether they have any interactions or not).
c. The number of unique users in the dataset. (excluding null values)
d. The number of user-article interactions in the dataset.
# a.
len(df['article_id'].unique())
# b.
print (len(set(df['article_id'].values)))
# same as the answer in project_test but I dont think this is correct
# unless only consider articles are in "Live" status and
# some of articles in "user-item-interactions.csv" dataset could be not in "Live" any more
print (len(set(df_content['article_id'].values)))
print (len(set.union(set(df['article_id'].values), set(df_content['article_id'].values))))
# c.
len(df['email'].unique()) # why it is 5148 in answer? because there are some observations have eail == NaN
# d.
df.shape[0]
unique_articles = len(df['article_id'].unique()) # The number of unique articles that have at least one interaction
total_articles = len(set(df_content['article_id'].values)) # The number of unique articles on the IBM platform
unique_users = len(df[df['email'].isnull() == False]['email'].unique()) # The number of unique users
user_article_interactions = df.shape[0] # The number of user-article interactions
4. Use the cells below to find the most viewed article_id, as well as how often it was viewed. After talking to the company leaders, the email_mapper function was deemed a reasonable way to map users to ids. There were a small number of null values, and it was found that all of these null values likely belonged to a single user (which is how they are stored using the function below).
most_viewed_article = (df
.groupby(['article_id'])
.count()['title']
.reset_index()
.sort_values(['title'], ascending=False)
.head(1)
)
most_viewed_article.columns = ['article_id', 'num_viewed']
most_viewed_article_id = str(most_viewed_article['article_id'].values[0]) # The most viewed article in the dataset as a string with one value following the decimal
max_views = most_viewed_article['num_viewed'].values[0] # The most viewed article in the dataset was viewed how many times?
## No need to change the code here - this will be helpful for later parts of the notebook
# Run this cell to map the user email to a user_id column and remove the email column
def email_mapper():
coded_dict = dict()
cter = 1
email_encoded = []
for val in df['email']:
if val not in coded_dict:
coded_dict[val] = cter
cter+=1
email_encoded.append(coded_dict[val])
return email_encoded
email_encoded = email_mapper()
del df['email']
df['user_id'] = email_encoded
# show header
df.head()
## If you stored all your results in the variable names above,
## you shouldn't need to change anything in this cell
sol_1_dict = {
'`50% of individuals have _____ or fewer interactions.`': median_val,
'`The total number of user-article interactions in the dataset is ______.`': user_article_interactions,
'`The maximum number of user-article interactions by any 1 user is ______.`': max_views_by_user,
'`The most viewed article in the dataset was viewed _____ times.`': max_views,
'`The article_id of the most viewed article is ______.`': most_viewed_article_id,
'`The number of unique articles that have at least 1 rating ______.`': unique_articles,
'`The number of unique users in the dataset is ______`': unique_users,
'`The number of unique articles on the IBM platform`': total_articles
}
# Test your dictionary against the solution
t.sol_1_test(sol_1_dict)
Unlike in the earlier lessons, we don't actually have ratings for whether a user liked an article or not. We only know that a user has interacted with an article. In these cases, the popularity of an article can really only be based on how often an article was interacted with.
1. Fill in the function below to return the n top articles ordered with most interactions as the top. Test your function using the tests below.
def get_top_articles(n, df=df):
'''
INPUT:
n - (int) the number of top articles to return
df - (pandas dataframe) df as defined at the top of the notebook
OUTPUT:
top_articles - (list) A list of the top 'n' article titles
'''
# Your code here
top_articles = (df
.groupby(['article_id', 'title'])
.count()['user_id']
.reset_index()
.sort_values(['user_id'], ascending=False)
.head(n)['title'])
return top_articles # Return the top article titles from df (not df_content)
def get_top_article_ids(n, df=df):
'''
INPUT:
n - (int) the number of top articles to return
df - (pandas dataframe) df as defined at the top of the notebook
OUTPUT:
top_articles - (list) A list of the top 'n' article titles
'''
# Your code here
top_articles = (df
.groupby(['article_id', 'title'])
.count()['user_id']
.reset_index()
.sort_values(['user_id'], ascending=False)
.head(n)['article_id'])
return top_articles # Return the top article ids
print(get_top_articles(10))
print(get_top_article_ids(10))
# Test your function by returning the top 5, 10, and 20 articles
top_5 = get_top_articles(5)
top_10 = get_top_articles(10)
top_20 = get_top_articles(20)
# Test each of your three lists from above
t.sol_2_test(get_top_articles)
1. Use the function below to reformat the df dataframe to be shaped with users as the rows and articles as the columns.
Use the tests to make sure the basic structure of your matrix matches what is expected by the solution.
df.head()
df[df['user_id'] == 1].sort_values(['article_id'])['article_id'].unique()
# create the user-article matrix with 1's and 0's
def create_user_item_matrix(df):
'''
INPUT:
df - pandas dataframe with article_id, title, user_id columns
OUTPUT:
user_item - user item matrix
Description:
Return a matrix with user ids as rows and article ids on the columns with 1 values where a user interacted with
an article and a 0 otherwise
'''
# Fill in the function here
df['time'] = 1
user_item = df.pivot_table(index='user_id',
columns='article_id',
values='time',
aggfunc='mean',
fill_value=0)
return user_item # return the user_item matrix
user_item = create_user_item_matrix(df)
## Tests: You should just need to run this cell. Don't change the code.
assert user_item.shape[0] == 5149, "Oops! The number of users in the user-article matrix doesn't look right."
assert user_item.shape[1] == 714, "Oops! The number of articles in the user-article matrix doesn't look right."
assert user_item.sum(axis=1)[1] == 36, "Oops! The number of articles seen by user 1 doesn't look right."
print("You have passed our quick tests! Please proceed!")
2. Complete the function below which should take a user_id and provide an ordered list of the most similar users to that user (from most similar to least similar). The returned result should not contain the provided user_id, as we know that each user is similar to him/herself. Because the results for each user here are binary, it (perhaps) makes sense to compute similarity as the dot product of two users.
Use the tests to test your function.
def find_similar_users(user_id, user_item=user_item):
'''
INPUT:
user_id - (int) a user_id
user_item - (pandas dataframe) matrix of users by articles:
1's when a user has interacted with an article, 0 otherwise
OUTPUT:
similar_users - (list) an ordered list where the closest users (largest dot product users)
are listed first
Description:
Computes the similarity of every pair of users based on the dot product
Returns an ordered
'''
# compute similarity of each user to the provided user
user_vector = user_item[user_item.index == user_id]
sim_vector = user_vector.dot(user_item.T).T
sim_vector.columns = ['product_dot_value']
# TDOD: one better solution is first reset index, then order the data frame and finally remove the own one
# sort by similarity
most_similar_users = sim_vector.sort_values(['product_dot_value'], ascending=False)
# create list of just the ids
most_similar_users = most_similar_users.reset_index()
# remove the own user's id
most_similar_users = most_similar_users[most_similar_users['user_id'] != user_id]
most_similar_users = most_similar_users \
.sort_values(['product_dot_value', 'user_id'], ascending=[False, True]) \
.reset_index()
most_similar_users = most_similar_users.rename(columns={"index": "original_index"})
return most_similar_users # return a list of the users in order from most to least similar
# Do a spot check of your function
print("The 10 most similar users to user 1 are: {}".format(find_similar_users(1)[:10]))
print("The 5 most similar users to user 3933 are: {}".format(find_similar_users(3933)[:5]))
print("The 3 most similar users to user 46 are: {}".format(find_similar_users(46)[:3]))
3. Now that you have a function that provides the most similar users to each user, you will want to use these users to find articles you can recommend. Complete the functions below to return the articles you would recommend to each user.
def get_article_names(article_ids, df=df):
'''
INPUT:
article_ids - (list) a list of article ids
df - (pandas dataframe) df as defined at the top of the notebook
OUTPUT:
article_names - (list) a list of article names associated with the list of article ids
(this is identified by the title column)
'''
# Your code here
df['article_id'] = df['article_id'].astype('string')
article_names = set(df[df['article_id'].isin(article_ids)]['title'])
return article_names # Return the article names associated with list of article ids
def get_user_articles(user_id, user_item=user_item):
'''
INPUT:
user_id - (int) a user id
user_item - (pandas dataframe) matrix of users by articles:
1's when a user has interacted with an article, 0 otherwise
OUTPUT:
article_ids - (list) a list of the article ids seen by the user
article_names - (list) a list of article names associated with the list of article ids
(this is identified by the doc_full_name column in df_content)
Description:
Provides a list of the article_ids and article titles that have been seen by a user
'''
# Your code here
user_item_df = user_item[user_item.index == user_id].T
user_item_df.columns = ['is_interacted']
article_ids = user_item_df[user_item_df['is_interacted'] == 1].index.values
article_ids = article_ids.astype(np.str_)
article_names = get_article_names(article_ids)
return article_ids, article_names # return the ids and names
def make_random_numbers(x):
# give a random number rows in one group
total = x['user_id'].count()
r = np.random.choice(range(99999), total, replace = False)
x['order_in_group'] = r
return x
def user_user_recs(user_id, m=10):
'''
INPUT:
user_id - (int) a user id
m - (int) the number of recommendations you want for the user
OUTPUT:
recs - (list) a list of recommendations for the user
Description:
Loops through the users based on closeness to the input user_id
For each user - finds articles the user hasn't seen before and provides them as recs
Does this until m recommendations are found
Notes:
Users who are the same closeness are chosen arbitrarily as the 'next' user
For the user where the number of recommended articles starts below m
and ends exceeding m, the last items are chosen arbitrarily
'''
# Your code here
sim_users = (find_similar_users(user_id)
.sort_values(['product_dot_value'], ascending=False)
.groupby('product_dot_value')
.apply(lambda x: make_random_numbers(x))
.sort_values(['product_dot_value', 'order_in_group'], ascending=False)
)
# filter articles that the target user has seen before - step 1
article_seen_ids, article_seen_names = get_user_articles(user_id)
article_seen_ids = list(article_seen_ids)
recs = []
for idx, row in sim_users.iterrows():
article_ids = []
user_id = row['user_id']
article_ids_tmp, article_names = get_user_articles(user_id)
# filter articles that the target user has seen before - step 2
article_ids_tmp = list(article_ids_tmp)
[article_ids.append(x) for x in article_ids_tmp if x not in article_seen_ids]
recs_size = len(recs)
ids_size = len(article_ids)
if (recs_size + ids_size > m):
'''
For the user where the number of recommended articles starts below m
and ends exceeding m, the last items are chosen arbitrarily
'''
temp_0 = article_ids[:(m-recs_size-1)] # 0-9
temp_1 = article_ids[(m-recs_size-1):] # 10
article_ids = temp_0 + [temp_1[np.random.randint(0, len(temp_1)-1)]]
for a_id in article_ids:
if (a_id not in recs):
recs.append(a_id)
if (len(recs) == m):
return recs
return recs # return your recommendations for this user_id
# Check Results
get_article_names(user_user_recs(1, 10)) # Return 10 recommendations for user 1
# Test your functions here - No need to change this code - just run this cell
assert set(get_article_names(['1024.0', '1176.0', '1305.0', '1314.0', '1422.0', '1427.0'])) == set(['using deep learning to reconstruct high-resolution audio', 'build a python app on the streaming analytics service', 'gosales transactions for naive bayes model', 'healthcare python streaming application demo', 'use r dataframes & ibm watson natural language understanding', 'use xgboost, scikit-learn & ibm watson machine learning apis']), "Oops! Your the get_article_names function doesn't work quite how we expect."
assert set(get_article_names(['1320.0', '232.0', '844.0'])) == set(['housing (2015): united states demographic measures','self-service data preparation with ibm data refinery','use the cloudant-spark connector in python notebook']), "Oops! Your the get_article_names function doesn't work quite how we expect."
assert set(get_user_articles(20)[0]) == set(['1320.0', '232.0', '844.0'])
assert set(get_user_articles(20)[1]) == set(['housing (2015): united states demographic measures', 'self-service data preparation with ibm data refinery','use the cloudant-spark connector in python notebook'])
assert set(get_user_articles(2)[0]) == set(['1024.0', '1176.0', '1305.0', '1314.0', '1422.0', '1427.0'])
assert set(get_user_articles(2)[1]) == set(['using deep learning to reconstruct high-resolution audio', 'build a python app on the streaming analytics service', 'gosales transactions for naive bayes model', 'healthcare python streaming application demo', 'use r dataframes & ibm watson natural language understanding', 'use xgboost, scikit-learn & ibm watson machine learning apis'])
print("If this is all you see, you passed all of our tests! Nice job!")
4. Now we are going to improve the consistency of the user_user_recs function from above.
def get_top_sorted_users(user_id, df=df, user_item=user_item):
'''
INPUT:
user_id - (int)
df - (pandas dataframe) df as defined at the top of the notebook
user_item - (pandas dataframe) matrix of users by articles:
1's when a user has interacted with an article, 0 otherwise
OUTPUT:
neighbors_df - (pandas dataframe) a dataframe with:
neighbor_id - is a neighbor user_id
similarity - measure of the similarity of each user to the provided user_id
num_interactions - the number of articles viewed by the user - if a u
Other Details - sort the neighbors_df by the similarity and then by number of interactions where
highest of each is higher in the dataframe
'''
# Your code here
user_item_interact_df = (df[df['user_id'] != user_id].groupby(['user_id'])['article_id'].count().reset_index())
user_similarity_df = find_similar_users(user_id)
neighbors_df = pd.merge(
user_item_interact_df,
user_similarity_df[['user_id', 'product_dot_value']],
on=['user_id'],
how='inner'
)
neighbors_df.columns = ['user_id', 'interact_num', 'similarity_value']
neighbors_df = neighbors_df.sort_values(['similarity_value', 'interact_num'], ascending=False)
return neighbors_df # Return the dataframe specified in the doc_string
################################################################################################
def get_user_top_articles(user_id, df=df):
top_articles_overall = (df
.groupby(['article_id', 'title'])
.count()['user_id']
.reset_index()
.sort_values(['user_id'], ascending=False)
)
top_articles_overall.columns = ['article_id', 'title', 'interact_num']
user_articles = df[df['user_id'] == user_id]['article_id']
user_articles = user_articles.drop_duplicates(keep='first')
top_articles = pd.merge(top_articles_overall, user_articles, on='article_id', how='inner') \
.sort_values(['interact_num'], ascending=False)
return top_articles
def user_user_recs_part2(user_id, m=10):
'''
INPUT:
user_id - (int) a user id
m - (int) the number of recommendations you want for the user
OUTPUT:
recs - (list) a list of recommendations for the user by article id
rec_names - (list) a list of recommendations for the user by article title
Description:
Loops through the users based on closeness to the input user_id
For each user - finds articles the user hasn't seen before and provides them as recs
Does this until m recommendations are found
Notes:
* Choose the users that have the most total article interactions
before choosing those with fewer article interactions.
* Choose articles with the articles with the most total interactions
before choosing those with fewer total interactions.
'''
# Your code here
sim_users = get_top_sorted_users(user_id)
# TODO; filter articles that the target user has seen before
recs = []
rec_names = []
for idx, row in sim_users.iterrows():
user_id = row['user_id']
articles = get_user_top_articles(user_id)
article_ids = list(articles['article_id'])
recs_size = len(recs)
ids_size = len(article_ids)
if (recs_size + ids_size > m):
article_ids = article_ids[:(m-recs_size)]
for a_id in article_ids:
if (a_id not in recs):
recs.append(a_id)
rec_names += articles[articles['article_id'] == a_id]['title'].values.tolist()
if (len(recs) == m):
return recs, rec_names
return recs, rec_names
# Quick spot check - don't change this code - just use it to test your functions
rec_ids, rec_names = user_user_recs_part2(20, 10)
print("The top 10 recommendations for user 20 are the following article ids:")
print(rec_ids)
print()
print("The top 10 recommendations for user 20 are the following article names:")
print(rec_names)
5. Use your functions from above to correctly fill in the solutions to the dictionary below. Then test your dictionary against the solution. Provide the code you need to answer each following the comments below.
### Tests with a dictionary of results
user1_most_sim = find_similar_users(1).loc[0, 'user_id'] # Find the user that is most similar to user 1
user131_10th_sim = find_similar_users(131).loc[9, 'user_id']# Find the 10th most similar user to user 131
## Dictionary Test Here
sol_5_dict = {
'The user that is most similar to user 1.': user1_most_sim,
'The user that is the 10th most similar to user 131': user131_10th_sim,
}
t.sol_5_test(sol_5_dict)
6. If we were given a new user, which of the above functions would you be able to use to make recommendations? Explain. Can you think of a better way we might make recommendations? Use the cell below to explain a better method for new users.
Provide your response here.
Since there is no any article interaction record that a new user has, it is somehow impossible to predict his taste of preference without any informative data about him. Therefore, articles with a higher interaction rankings (i.e., get_top_articles function) are much highly recommended for that new user.
One better solution is that:
7. Using your existing functions, provide the top 10 recommended articles you would provide for the a new user below. You can test your function against our thoughts to make sure we are all on the same page with how we might make a recommendation.
new_user = '0.0'
# What would your recommendations be for this new user '0.0'? As a new user, they have no observed articles.
# Provide a list of the top 10 article ids you would give to
# To a new user, recommending articles with top interactions (as the most popular ones on the platform)
top_articles = (df
.groupby(['article_id', 'title'])
.count()['user_id']
.reset_index()
.sort_values(['user_id'], ascending=False)
.head(10)
)
new_user_recs = set(top_articles['article_id'])# Your recommendations here
assert set(new_user_recs) == set(['1314.0','1429.0','1293.0','1427.0','1162.0','1364.0','1304.0','1170.0','1431.0','1330.0']), "Oops! It makes sense that in this case we would want to recommend the most popular articles, because we don't know anything about these users."
print("That's right! Nice job!")
Another method we might use to make recommendations is to perform a ranking of the highest ranked articles associated with some term. You might consider content to be the doc_body, doc_description, or doc_full_name. There isn't one way to create a content based recommendation, especially considering that each of these columns hold content related information.
1. Use the function body below to create a content based recommender. Since there isn't one right answer for this recommendation tactic, no test functions are provided. Feel free to change the function inputs if you decide you want to try a method that requires more input values. The input values are currently set with one idea in mind that you may use to make content based recommendations. One additional idea is that you might want to choose the most popular recommendations that meet your 'content criteria', but again, there is a lot of flexibility in how you might make these recommendations.
import re
import nltk
from nltk.corpus import stopwords
from nltk.tokenize import word_tokenize
from nltk.stem import WordNetLemmatizer, PorterStemmer
nltk.download(['punkt', 'wordnet', 'stopwords'])
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.base import BaseEstimator, TransformerMixin
df.head()
df_content.head()
def tokenize(text):
# 1. remove url and replace url string as 'urlplaceholder'
url_regex = 'http[s]?://(?:[a-zA-Z]|[0-9]|[$-_@.&+]|[!*\(\),]|(?:%[0-9a-fA-F][0-9a-fA-F]))+'
detected_urls = re.findall(url_regex, text)
for url in detected_urls:
text = text.replace(url, "urlplaceholder")
# 2. remove punctuation
text = re.sub(r"[^a-zA-Z0-9]"," ",text)
# 3. work tokenization
tokens = word_tokenize(text)
# 4. remove stop words
tokens = [tok for tok in tokens if tok not in stopwords.words("english")]
lemmatizer = WordNetLemmatizer() # lemmatization method
clean_tokens = []
for tok in tokens:
# 3. converting lowercase and removing space in tokens
clean_tok = lemmatizer.lemmatize(tok).lower().strip()
clean_tokens.append(clean_tok)
return clean_tokens
def insert_title(title, doc_full_name):
if (title is np.nan):
title = doc_full_name
return title
def generate_tfidf_mat(df=df, df_content=df_content):
'''
Generate matrix (in a pandas dataframe format) of articles by tokens (i.e., title)
'''
pipeline = Pipeline([
('vect', CountVectorizer(tokenizer=tokenize)),
('tfidf', TfidfTransformer()),
])
# clean df_content
df_content = df_content.drop_duplicates(subset=['article_id'], keep='first')
df_content['article_id'] = df_content['article_id'].apply(lambda x: str(x)+'.0')
df_content.index.name = None
df = df.drop_duplicates(subset=['article_id'], keep='first')
df_cleansed = pd.merge(df, df_content, on='article_id', how='outer')
df_cleansed['title'] = df_cleansed.apply(lambda row: insert_title(row['title'], row['doc_full_name']), axis=1)
df_cleansed.set_index(df_cleansed['article_id'], inplace=True)
df_cleansed.index.name = None
df_tfidf_mat = pipeline.fit_transform(df_cleansed['title'])
article_token_df = pd.DataFrame.sparse.from_spmatrix(df_tfidf_mat) # row: article id; col: token id
article_token_df.set_index(df_cleansed['article_id'], inplace=True)
article_token_df.index.name = None
return df_cleansed, article_token_df
df_cleansed, article_token_df = generate_tfidf_mat()
df_cleansed.head()
article_token_df
def find_similar_articles(article_id, article_token_df=article_token_df):
'''
INPUT:
article_id - (int) a article_id
article_token_df - (pandas dataframe) matrix of articles by tokens (i.e., title)
OUTPUT:
most_similar_articles - (list) an ordered list where the closest articles (largest dot product articles)
are listed first
Description:
Computes the similarity of every pair of articles based on the dot product
Returns an ordered list of articles with a higher similarity value
'''
# compute similarity of each article to the provided article
article_vector = article_token_df[article_token_df.index == article_id]
sim_vector = article_vector.dot(article_token_df.T).T
sim_vector.columns = ['product_dot_value']
# create list of just the ids
sim_vector['article_id'] = sim_vector.index
# remove the own user's id
most_similar_articles = sim_vector[sim_vector['article_id'] != article_id]
# sort by similarity
most_similar_articles = most_similar_articles \
.sort_values(['product_dot_value', 'article_id'], ascending=[False, True])
return most_similar_articles # return a list of the most_similar_articles in order from most to least similar
df_content.head()
def make_content_recs(user_id, n=10, df=df, df_content=df_content, df_cleansed=df_cleansed):
'''
INPUT:
user_id - (int) a user_id
m - (int) the number of recommendations you want for the user
df - pandas dataframe with article_id, title, user_id columns
df_content - pandas dataframe with doc_body, doc_description, doc_full_name, doc_status, article_id
df_cleansed - pandas dataframe with article_id, title, user_id, time, doc_body, doc_description, doc_full_name, doc_status;
merged from df and df_content
OUTPUT:
recs_id - (list) a list of recommendations for the user by article id
recs_name - list) a list of recommendations for the user by article title
Description:
Makes content based recommendations for a specific input user_id
For each user - finds articles the user hasn't seen before and provides them as recs
Does this until n recommendations are found
'''
# 1. Get article ids that user has already seen; the list ordered by interaction rank
article_seen_ids = list(df[df['user_id'] == user_id]['article_id'].values)
# Get the user-article interaction rank
interact_ranked_df = (df
.groupby(['article_id', 'title'])
.count()['user_id']
.reset_index()
.sort_values(['user_id'], ascending=False))
interact_ranked_df.rename(columns={'user_id': 'interact_num'}, inplace=True)
article_seen_ids = interact_ranked_df[interact_ranked_df['article_id'].isin(article_seen_ids)]['article_id']
# article_seen_ids = ['0.0', '2.0']
# Get a list of articles are in 'Live' status
articles_in_live = [str(art_id)+'.0' for art_id in df_content['article_id'].unique()]
# 2. Make recommendation based on articles already seen
recs_id = []
recs_name = []
candidate_articles = []
# Loops => get recommendation based on seen articles with a similar content
for art_id in article_seen_ids:
# Get articles similar to the target article ranked by product dot value (i.e., similarity value)
most_similar_articles_all = find_similar_articles(art_id)
# Filter top n articles still in "Live" status
most_similar_articles = most_similar_articles_all[~most_similar_articles_all['article_id']
.isin(articles_in_live)].head(n).reset_index()
# Get top n articles names
most_similar_articles['article_name'] = list(set(df_cleansed[df_cleansed['article_id']
.isin(most_similar_articles['article_id'])]['title']))
# The most similar article directly added to the recommendations list
recs_id.append(most_similar_articles.loc[0, 'article_id'])
recs_name.append(most_similar_articles.loc[0, 'article_name'])
if (len(recs_id) == n):
return (recs_id, recs_name)
# The rest similar articles added to the candicates list
for idx, row in most_similar_articles.loc[1:, ['article_id', 'article_name']].iterrows():
candidate_articles.append(list(row.values))
# 3. For the rest, randomly select similar articles from the candidates
candidate_articles_df = pd.DataFrame(data = candidate_articles, columns = ['article_id', 'article_name'])
candidate_articles_df.drop_duplicates(subset=['article_id'], keep='first')
recs_size = len(recs_id)
sample_size = recs_size if recs_size < (n-recs_size) else (n-recs_size)
# Randomly select a specified number of rows
recs_candicate_articles = candidate_articles_df.sample(n=sample_size)
# Add articles to the recommendations list
for idx, row in recs_candicate_articles.iterrows():
recs_id.append(row['article_id'])
recs_name.append(row['article_name'])
return (recs_id, recs_name)
make_content_recs(user_id=131, n=10)
2. Now that you have put together your content-based recommendation system, use the cell below to write a summary explaining how your content based recommender works. Do you see any possible improvements that could be made to your function? Is there anything novel about your content based recommender?
Write an explanation of your content based recommendation system here.
The process of my content based recommender executes:
3. Use your content-recommendation system to make recommendations for the below scenarios based on the comments. Again no tests are provided here, because there isn't one right answer that could be used to find these content based recommendations.
# make recommendations for a brand new user
print("recommendations for a brand new user:\n")
print(list(get_top_article_ids(10, df=df)))
print(list(get_top_articles(10, df=df)))
print ('############################################################################################################ \n')
# make a recommendations for a user who only has interacted with article id '1427.0'
print("Recommendation for a user who only has interacted with article id '1427.0':\n")
similar_articles_ids = find_similar_articles(article_id='1427.0').head(10)['article_id']
print (list(similar_articles_ids))
print (list(set(df_cleansed[df_cleansed['article_id'].isin(similar_articles_ids)]['title'])))
In this part of the notebook, you will build use matrix factorization to make article recommendations to the users on the IBM Watson Studio platform.
1. You should have already created a user_item matrix above in question 1 of Part III above. This first question here will just require that you run the cells to get things set up for the rest of Part V of the notebook.
# Load the matrix here
user_item_matrix = pd.read_pickle('../test/user_item_matrix.p')
# quick look at the matrix
user_item_matrix.head()
2. In this situation, you can use Singular Value Decomposition from numpy on the user-item matrix. Use the cell to perform SVD, and explain why this is different than in the lesson.
# Perform SVD on the User-Item Matrix Here
u, s, vt = np.linalg.svd(user_item_matrix.values, full_matrices=True) # use the built in to get the three matrices
print (u.shape, vt.shape, s.shape)
Provide your response here.
The data provided in the lessson having data points in null (i.e., missing values), which can not be factorised by a pure SVD method. FunkSVD is an alternative technique that works works well with matric having missing values. However, the user-item matrix generated in this exercies has been initially cleansed as 1's when a user has interacted with an article, 0 otherwise. That is the main reason why here, we can perfrom SVD instead.
3. Now for the tricky part, how do we choose the number of latent features to use? Running the below cell, you can see that as the number of latent features increases, we obtain a lower error rate on making predictions for the 1 and 0 values in the user-item matrix. Run the cell below to get an idea of how the accuracy improves as we increase the number of latent features.
num_latent_feats = np.arange(10,700+10,20)
sum_errs = []
for k in num_latent_feats:
# restructure with k latent features
s_new, u_new, vt_new = np.diag(s[:k]), u[:, :k], vt[:k, :]
# take dot product
user_item_est = np.around(np.dot(np.dot(u_new, s_new), vt_new))
# compute error for each prediction to actual value
diffs = np.subtract(user_item_matrix, user_item_est)
# total errors and keep track of them
err = np.sum(np.sum(np.abs(diffs)))
sum_errs.append(err)
plt.plot(num_latent_feats, 1 - np.array(sum_errs)/df.shape[0]);
plt.xlabel('Number of Latent Features');
plt.ylabel('Accuracy');
plt.title('Accuracy vs. Number of Latent Features');
4. From the above, we can't really be sure how many features to use, because simply having a better way to predict the 1's and 0's of the matrix doesn't exactly give us an indication of if we are able to make good recommendations. Instead, we might split our dataset into a training and test set of data, as shown in the cell below.
Use the code from question 3 to understand the impact on accuracy of the training and test sets of data with different numbers of latent features. Using the split below:
df_train = df.head(40000)
df_test = df.tail(5993)
def create_test_and_train_user_item(df_train, df_test):
'''
INPUT:
df_train - training dataframe
df_test - test dataframe
OUTPUT:
user_item_train - a user-item matrix of the training dataframe
(unique users for each row and unique articles for each column)
user_item_test - a user-item matrix of the testing dataframe
(unique users for each row and unique articles for each column)
test_idx - all of the test user ids
test_arts - all of the test article ids
'''
# Your code here
user_item_train = create_user_item_matrix(df_train)
user_item_test = create_user_item_matrix(df_test)
test_idx = user_item_test.index
test_arts = user_item_test.columns
return user_item_train, user_item_test, test_idx, test_arts
user_item_train, user_item_test, test_idx, test_arts = create_test_and_train_user_item(df_train, df_test)
# 4a
user_item_test.index.isin(user_item_train.index).sum()
# 4b -> cold start problem means "new" users to a platform dont have any "records" (e.g., ratings, interactions...)
len(test_idx) - 20
# 4c
user_item_test.columns.isin(user_item_train.columns).sum()
# 4d
len(test_arts) - 574
# Replace the values in the dictionary below
a = 662
b = 574
c = 20
d = 0
sol_4_dict = {
'How many users can we make predictions for in the test set?': c,
'How many users in the test set are we not able to make predictions for because of the cold start problem?': a,
'How many articles can we make predictions for in the test set?': b,
'How many articles in the test set are we not able to make predictions for because of the cold start problem?': d
}
t.sol_4_test(sol_4_dict)
5. Now use the user_item_train dataset from above to find U, S, and V transpose using SVD. Then find the subset of rows in the user_item_test dataset that you can predict using this matrix decomposition with different numbers of latent features to see how many features makes sense to keep based on the accuracy on the test data. This will require combining what was done in questions 2 - 4.
Use the cells below to explore how well SVD works towards making predictions for recommendations on the test data.
# fit SVD on the user_item_train matrix
u_train, s_train, vt_train = np.linalg.svd(user_item_train, full_matrices=True) # fit svd similar to above then use the cells below
s_train.shape
# Use these cells to see how well you can use the training
# decomposition to predict on test data
from sklearn.metrics import accuracy_score
def make_prediction(u, s, vt):
'''
Calculates SVD values with u, sigma and v (transpose) matices
'''
pred = np.round(np.dot(np.dot(u, s), vt))
return pred
def predict_interaction(n, u_train, s_train, vt_train, user_rows, article_rows):
'''
Makes predictions from training set to test set with SVD
'''
# restructure with k latent features
u_train_lat, s_train_lat, vt_train_lat = u_train[:, :n], np.diag(s_train[:n]), vt_train[:n, :]
u_test_lat, vt_test_lat = user_rows[:, :n], article_rows[:n, :]
train_preds = make_prediction(u_train_lat, s_train_lat, vt_train_lat)
test_preds = make_prediction(u_test_lat, s_train_lat, vt_test_lat)
return train_preds, test_preds
def generate_matrix(user_item_train, user_item_test, u_train, s_train, vt_train):
'''
Generates the training set errors and test set errors matrix
'''
user_idx = user_item_train.index.isin(user_item_test.index)
article_idx = user_item_train.columns.isin(user_item_test.columns)
user_item_test2 = user_item_test.loc[user_item_train.index[user_idx==True], user_item_train.columns[article_idx==True]]
user_rows = u_train[user_idx, :]
article_rows = vt_train[:, article_idx]
# print (user_rows.shape, article_rows.shape)
train_errs = []
test_errs = []
for n_features in np.arange(0, 720, 20):
train_preds, test_preds = predict_interaction(n_features, u_train, s_train, vt_train, user_rows, article_rows)
# compute prediction accuracy
train_errs.append(accuracy_score(user_item_train.values.flatten(), train_preds.flatten()))
test_errs.append(accuracy_score(user_item_test2.values.flatten(), test_preds.flatten()))
return train_errs, test_errs
train_errs, test_errs = generate_matrix(user_item_train, user_item_test, u_train, s_train, vt_train)
def plot_train_test_error(train_errs, test_errs):
'''
Plots the trend of accuracy on training and test set with different number of latern features
'''
plt.figure()
plt.plot(np.arange(0, 720, 20), train_errs, label='Train')
plt.plot(np.arange(0, 720, 20), test_errs, label='Test')
plt.xlabel('Number of Latent Features')
plt.ylabel('Accuracy')
plt.title('Accuracy Testing on Train and Test Set with different Number of Latent Features')
plt.legend()
plt.show()
plot_train_test_error(train_errs, test_errs)
# Test with selecting training/testing set randomly
df_train_2 = df.sample(frac = 0.75, random_state=200)
df_test_2 = df.drop(df_train_2.index)
user_item_train_2, user_item_test_2, test_idx_2, test_arts_2 = create_test_and_train_user_item(df_train_2, df_test_2)
u_train_2, s_train_2, vt_train_2 = np.linalg.svd(user_item_train_2, full_matrices=True)
train_errs_2, test_errs_2 = generate_matrix(user_item_train_2, user_item_test_2, u_train_2, s_train_2, vt_train_2)
plot_train_test_error(train_errs=train_errs_2, test_errs=test_errs_2)
def explained_variance(sigma, n_components):
"""
Computes explained variance number of components
"""
# explained variance
total_var = np.sum(sigma**2)
var_exp = np.sum([np.square(i) for i in sigma[:n_components]])
perc_exp = (var_exp / total_var) * 100
return round(perc_exp, 4)
explained_variance(sigma=s_train, n_components=300)
explained_variance(sigma=s_train_2, n_components=300)
6. Use the cell below to comment on the results you found in the previous question. Given the circumstances of your results, discuss what you might do to determine if the recommendations you make with any of the above recommendation systems are an improvement to how users currently find articles?
Your response here.
From the cells above, two types of approach in training/test datasets splitting have been used. The first one is having 40000 observations from the top as the traininng set, and the rest of them are the test set. In comparison, the second one is having sample function to randomly subset rows into training or test sets. The other operations in the both approaches are the same. Finally, we have created two plots to compare the training and test errors.
From the plots we can find that both figure has a similar circumstance as the the number latent feature rises up, the accuracy of traning set will also increase gradually and reach the acrrucay score at 1.000. However, the performance on test set will decrease till about 0.965 and 0.985 respectively.
One of the reason for this kind of circumstance is due to the imbalance of common users and articles distribution in training and test datasets. As the number of latent features increases, the model becomes more and more overfitted on the traing set and makes predictions on the test set more negatively with lower accuracy.
The are several approaches can be used to make improvements on the recommendation engine such as scaling up the datasets with more user-article interaction pair records, A/B testing in changing the way of article rankings, asking feedbacks from online users (on testing environment, not on production environment) for further evalution of how closer the recommendation system can get the "taste" of readers.
Using your workbook, you could now save your recommendations for each user, develop a class to make new predictions and update your results, and make a flask app to deploy your results. These tasks are beyond what is required for this project. However, from what you learned in the lessons, you certainly capable of taking these tasks on to improve upon your work here!
Congratulations! You have reached the end of the Recommendations with IBM project!
Tip: Once you are satisfied with your work here, check over your report to make sure that it is satisfies all the areas of the rubric. You should also probably remove all of the "Tips" like this one so that the presentation is as polished as possible.
Before you submit your project, you need to create a .html or .pdf version of this notebook in the workspace here. To do that, run the code cell below. If it worked correctly, you should get a return code of 0, and you should see the generated .html file in the workspace directory (click on the orange Jupyter icon in the upper left).
Alternatively, you can download this report as .html via the File > Download as submenu, and then manually upload it into the workspace directory by clicking on the orange Jupyter icon in the upper left, then using the Upload button.
Once you've done this, you can submit your project by clicking on the "Submit Project" button in the lower right here. This will create and submit a zip file with this .ipynb doc and the .html or .pdf version you created. Congratulations!
from subprocess import call
call(['python', '-m', 'nbconvert', 'Recommendations_with_IBM.ipynb'])